RobotWorld benchmark tests agents across 84 simulated robot tasks

What the research found

RobotWorld is a 84-task simulation benchmark for general agents controlling robots, finding capabilities fail to compose into reliable behavior; Astra and Opus 5.5 show different strengths.

Claims and how well the evidence supports them

1. RobotWorld contains 84 tasks spanning manipulation, mobile manipulation, locomotion, driving and aerial control 99%
2. Astra succeeds more often on spatial and constrained-contact goals while Opus 5.5 succeeds more on continuous-balance goals 97%
3. The paper was submitted to arXiv on 7 October 2026 as v1 99%

Evidence

• arxiv.org https://arxiv.org/abs/2610.10409

Actionability rated none.